GPT-Red – A Red Teamer to Find Prompt Injection Vulnerabilities in GPT 5.6 Sol

OpenAI developed GPT-Red, an internal automated red-teaming model trained via self-play reinforcement learning to generate and iterate prompt-injection attacks against AI models; testing showed high success rates against earlier models (and better performance than human red teamers) and revealed real-world risks like data exfiltration and unauthorized agent actions, with findings used to harden GPT-5.6 and mitigation controls implemented to keep offensive capabilities segregated from public models.